Realistic tropical cyclone wind and pressure fields can be reconstructed from sparse data using deep learning

Tropical cyclones are responsible for large-scale loss of life and property ^1–4, motivating accurate risk assessment and forecasting. These objectives require accurate reconstructions of storms’ wind and pressure fields which assimilate real-time observations ^5–9, but current methods used for these reconstructions remain computationally expensive and limited ¹⁰. Here, we show that a physics-informed neural network ^11,12 can be a promising and computationally efficient algorithm for tropical cyclone data assimilation. Using synthetic training data sparsely sampled from hurricanes simulated in a forecast model, a physics-informed neural network is able to reconstruct full realistic 2- and 3-dimensional wind and pressure fields which capture key features of the cyclone. We also demonstrate how a set of sparse, real-time observations, can be used to accurately reconstruct Hurricane Ida. Our results highlight how recent advances in deep learning can augment data assimilation schemes. The methods are also general and can be applied to other flow problems.